Databases such as the Cochrane Central Register of Controlled Trials (CENTRAL), PubMed, Embase, Web of Science, China National Knowledge Infrastructure (CNKI), VIP information database and Wanfang Data were searched from inception to January 31, 2021. We also checked reference lists and conference proceedings manually to obtain additional relevant data. No language or publication date restrictions. The details of the search strategy in PubMed are shown in Supplemental Table 1 .

Studies which met the following inclusion criteria were included: (1) parallel controlled RCTs, (2) evaluating the effects of diet on fertility in women with a clear diagnosis of PCOS, such as the Rotterdam Consensus, the National Institutes of Health (NIH) diagnostic criteria, Androgen Excess and PCOS Society (AE–PCOS) Position Statement, as well as the China Medical Association diagnostic criteria (consisted of the following two aspects: a. Suspected PCOS: oligomenorrhoea or amenorrhoea or abnormal uterine bleeding is required, meanwhile accompanied by evidence of clinical or biochemical hyperandrogenism and/or PCOM; b. Confirmed PCOS: based on the suspected PCOS diagnosis conditions, it can be diagnosed by excluding other diseases that can cause hyperandrogenism), (3) studies with exercise/medication as a cointervention in both intervention/control arms were also considered. The primary outcomes referred to clinical pregnancy rate (defined as the presentence of intrauterine gestational sac with foetal heartbeats), miscarriage rate (defined as loss of intrauterine pregnancy before 20 completed weeks of gestation) and ovulation rate (determined by ultrasound or increased progesterone) (26); the secondary outcomes included menstrual regularity rate, Anti-Müllerian Hormone (AMH), free androgen index (FAI), sex hormone-binding globulin (SHBG), total testosterone (T), Ferriman-Gallwey score. All outcomes were measured before and after the observation.

The exclusion criteria were as follows: (1) quasi-randomized trials, cohort or case-control studies, reviews, meta-analyses, case reports, animal or cell experiments, (2) women with other causes for hyperandrogenism and abnormal ovulation, or any cardiovascular and cerebrovascular diseases, psychiatric, or neurological problems, (3) interventions referring to single dietary components (e.g., vitamins, calcium), and (4) studies with insufficient data and unreported target outcomes.

Titles and abstracts of all potential studies were scanned to eliminate duplicated and ineligible studies. When the information was inadequate to make a decision, we sought further details from the original authors. Any discrepancies were resolved by discussion or consensus with the corresponding author.

Two authors extracted the predefined information from eligible studies independently. Additional information was further sought from the trials which appeared to be eligible but with unclear explanation of methodology or unsuitable data for meta-analysis. We cross-checked the data to minimize potential errors, and solved disagreements by discussion with the corresponding author. Information was collected from the included trials regarding the following aspects: (1) study characteristics, including first author, year of publication and location, (2) participants, including sample size and diagnostic criteria for PCOS, (3) interventions, including dietary protocols, frequency and duration of treatment, and (4) outcome data at baseline and follow-up.

Two authors assessed the risk of bias by the Cochrane Collaboration’s tool in eligible trials. Studies were deemed as low, unclear risk, or high bias based on the following domains: selection bias, performance bias, detection bias, attrition bias, reporting bias and other bias (27).

Review Manager 5.4 were applied for statistical analysis according to the Cochrane Handbook for Systematic Reviews of Interventions (28). P < 0.05 represented statistical significance. Results reported as binary variables were expressed as risk ratio (RR) with 95% confidence intervals. Data represented in continuous forms were pooled for meta-analysis as the mean difference (MD) with 95% confidence intervals if all studies reported the same scales. When data were reported on different methods or scales, the standardized mean difference (SMD) with 95% confidence intervals was calculated.

Due to inevitable clinical heterogeneity between studies, random-effects model was a more appropriate method to calculate summary effect measures. Statistical heterogeneity within comparisons was evaluated by Cochran’s Q test and quantified by the I-squared (I² ) statistic. I² values < 40% might be important, 30%-60%were deemed moderate, 50% to 90% were deemed substantial, and 75% to 100% were deemed considerable heterogeneity (27). Regarding missing data which was unobtainable from the original investigators, the analysis was performed on an intention-to-treat (ITT) basis for primary outcomes (i.e. including all participants in analysis, in the groups to which they were randomised). All imputation were subjected to sensitivity analysis. We also conducted meta-regression and subgroup analyses to explain potential sources of heterogeneity between studies. Subgroup analyses based on the predefined factors, such as dietary patterns, intervention duration (< 3 months, 3-6 months or > 6 months), diagnostic criteria and calorie restriction were also conducted to explore mediator effects of dietary modification.

Sensitivity analyses were performed to examine the robustness of pooled estimates. When there were more than ten trials included in the analysis, the potential publication bias was investigated by Egger’s test and visual inspection of funnel plots.

1123 studies were identified by the preliminary search. 456 records were removed due to duplication, and 667 studies were remained for further scanning of titles and abstracts. Among them, 479 items were excluded. We retrieved 188 full-text articles for detailed evaluation, and 168 trials were excluded for not meeting the inclusion criteria. Finally, 20 RCTs were included for meta-analysis. Details were shown in a PRISMA flow diagram ( Figure 1 ).

The general characteristics of the included studies are outlined in Table 1 . Except for the trial conducted by Gower et al. (crossover study) (34), all of them were parallel-design and single-canter RCTs conducted in China (40, 42–47), Iran (35, 36, 39, 41), the United States (30, 34), Australia (29, 32), the United Kingdom (31), Canada (48), Denmark (33), Egypt (37) and Mexico (38) between 2003 and 2020. The diagnosis of PCOS in the analysis could be divided into four categories: twelve trials under the Rotterdam Consensus (31, 33, 35, 36, 38–41, 43, 44, 46, 47), six trials following the NIH diagnostic criteria (29, 30, 32, 34, 37, 45), and the remaining two confirmed by the AE–PCOS (48) and China Medical Association (42) diagnostic criteria, respectively. Regarding dietary patterns, nine trials evaluated the low-carbohydrate diet (29, 30, 32, 33, 40, 42, 43, 46, 47); six trials respectively the low glycemic index/load diet (LGI/LGL) (31, 34, 37, 38, 45, 48); four trials evaluated the DASH diet (35, 36, 39, 41); and one trial evaluated the Mediterranean diet (44). The duration of diet ranged from one month to one year. Most of them had a medium duration (3-6 months), while four trials were within two months (30, 34–36) and one lasted for one year (42).

Fourteen studies provided data on randomization methods (29–33, 35–37, 39, 41, 43, 46–48), with two explaining the allocation concealment (41, 47). Blinding was only performed in four trials (35, 36, 39, 41), and were considered as low risk of bias. Four trials with participant-reported outcomes were judged as high risk, as the lack of participant blinding might introduce bias in these studies (29, 42–44). Five trials analyzed data on the ITT principle and were deemed as low risk (33, 35, 36, 41, 48). Five trials mentioning trial registration (35, 36, 39, 41, 48) were considered as low risk of reporting bias ( Figure 2 ) (49).

Meta-regression analyses were available only for clinical pregnancy rate. Factors, such as dietary patterns (regression coefficient β = 0.831; SE = 0.649; P = 0.241), treatment duration (regression coefficient β = 1.780; SE = 0.987; P = 0.114), diagnostic criteria (regression coefficient β = -0.937; SE = 0.908; P = 0.336) and calorie restriction (regression coefficient β = 0.726; SE = 0.591; P = 0.260) had no significant association with the study effect size. Meta-regression analyses were attempted to explain the heterogeneity among the studies, but inferences were limited by the paucity of available studies.

We conducted sensitivity analyses by restricting studies to studies without a high risk of bias. When excluding trials deemed as high risk of bias, the overall estimates remained unchanged, indicating the majority of conclusions were stable and not affected by the low-quality trials. We also performed sensitivity analysis the imputation of primary outcomes. Similarly, when compared the pooled estimates of imputation and available data, no difference was noted. Given the limited number of studies (< 10), Egger’s test and the forest plot can be low-powered. Thus, we could only conduct tests on clinical pregnancy rate. The P-value of Egger’s test was 0.931, indicating no evidence of publication bias in this outcome. The funnel plot did not show major asymmetries ( Figure 5 ).

In this systematic review and meta-analysis, the pooled data of 20 RCTs (1113 participants) showed that diet was not only associated with significantly improved fertility, but also mitigated hyperandrogenism in women with PCOS, which reiterated and extended those of previous reviews about the role of diet in endocrine, anthropometry and metabolism. In addition, these effects were associated with dietary patterns and treatment duration.

From the results of subgroup analyses, we found that low-carbohydrate diets tended to be better on improving pregnancy rate, reducing the risk of miscarriage and optimizing ovulation function. Our findings supported other notion in this topic. Several reports to date have showed that high-carbohydrate diets with a high glycemic index were associated with the increased risk of infertility concerning ovulatory disorders in apparently healthy women, while reducing carbohydrate consumption could influence the fertility and ovulatory function in turn (13, 50). Recently, there was evidence that the type of carbohydrate intake, such as low-glycemic index/load (LGI/LGL) food, was more important than the total amount received (51, 52). However, due to limited number of articles investigating the effectiveness of LGI/LGD diets on reproductive outcomes among women with PCOS, we were uncertain about its role in PCOS population

Consistent with previous research, we also found that calorie-restricted diets might be more salient in hyperandrogenism based on the subgroup analyses (53, 54). Hypocaloric diets could not only improve insulin sensitivity and regulate glycometabolism (55–57), but also advantageous for eliciting fast and significant weight loss, which exhibits a critical role in ameliorating PCOS phenotype. Weight reduction induced by calorie restriction is associated with reduced fat mass and preserved lean body mass (58), thus increasing the production of SHBG by the liver and reducing the levels of free testosterone (59, 60). However, dietary with energy limitation showed no effects in ovulation rate, and the improvement of clinical pregnancy rate, menstrual regularity rate and AMH level in women without calorie restriction were more obvious than those intaking fewer calories, which indicated that the benefits of diet might not just depend on weight loss, as not all PCOS patients with IR are overweight or obese and a higher incidence of IR have been reported in PCOS with normal weight (61, 62), suggesting that dietary management ought to go beyond weight loss. Of note, follicular development and ovulation require energy and energy requirements change during the menstrual cycle (63–65). Hence, it would be simplistic to claim for a beneficial effect of calorie restriction in all circumstances, since calorie restriction and consequent negative energy balance can also be harmful. Given this, different menstrual periods should also be considered during the calorie limitation. In our research, the favorable effects might also be associated with the treatment duration, as revealed in the subgroup analyses that the longer the duration, the greater the improvement was. Therefore, the diet treatment should be long term, dynamic and adapted to the changing circumstances, personal needs and expectations of the individual patient.

In our research, diet interventions were proved to increase the rate of decline of AMH, a well-recognized biomarker of ovarian reserve. Serum AMH concentration is higher in women with PCOS than in healthy women, which is related to severity of hyperandrogenism and oligo-anovulation (66, 67). A number of studies have reported that excess AMH could slow down initial follicular growth, decrease apoptosis of granulosa cells in small follicles with an anti-atretic effect, and cause follicular arrest in large antral follicles (68–70). Additionally, there is also a hypothesis that AMH appears to be able to exert its action at the hypothalamus and the pituitary level, which could either be at the origin of, or contribute to, the vicious circle of neuroendocrine and gonadal dysregulation encountered in PCOS (68). Therefore, the declined AMH levels might not only account for the reduced follicle excess of PCOM, but also line up with the elevated ovulation rate and ameliorative hyperandrogenism, thus improving the fertility outcomes.

The criteria used to diagnose PCOS were not uniform in this review, which might result in further clinical heterogeneity between studies. It has been reported that the overall prevalence of PCOS according to NIH criteria is 6%, while the pooled prevalence is 10% when applying the Rotterdam or AE-PCOS Society criteria. Studies in accordance with the NIH criteria, might narrow the phenotypic spectrum of PCOS and, thus limiting real PCOS population, as the morphology of polycystic ovarian is not considered as a diagnostic feature (71). However, the higher pooled prevalence estimates with the Rotterdam and AE-PCOS criteria is attributed to the inclusion of ovarian morphology and the ultrasound examination may provide false positive reports of PCOM (72). Besides, both oligo anovulation and PCOM are common in adolescent girls. Given this, the prevalence estimate might be exaggerated based on the Rotterdam criteria and people not suffering from PCOS truly might also be included. Therefore, due to the uncertainty surrounding the diagnosis of PCOS, and relative dearth of studies, we were unable to make conclusions concerning different diagnosis criteria.

Our research has unique strengths. To the best of our knowledge, this study is a frontier analysis to evaluate the role of diet on reproductive health in women with PCOS, in order to provide appropriate nutrition advice for clinical practice. We also performed detailed subgroup analysis based on different dietary patterns, treatment durations, diagnostic criteria, and whether energy was restricted, which may have significant impacts on the results. Since our research has been registered on PROSPERO, all the procedures were faithfully executed accordingly, as well as rigorous inclusion criteria, thus enhancing our results with more credibility and validity.

However, there were several limitations to be taken into consideration. Due to lack of livebirth rate, studies included were insufficient to address the role of diet on reproductive health comprehensively. Additionally, the evidence involved few countries and ethnic groups, which made the results difficult to be generalize. Besides, given the limited number of trials and small sample size in certain outcomes, the findings might be insufficient to ensure a significant difference.

More well-designed studies are warranted to confirm the effects of dietary intervention on reproductive health in PCOS population. PCOS is a heterogeneous condition with different phenotypes. However, no included trials targeted a specific phenotype, which made the results difficult to generalize. Future work should focus on the relationship between reproductive health in particular phenotype and dietary intervention, thus investigating the effects accordingly. Sociodemographic disparities may have an impact on the effects of diet, such as economic status and educational attainment. Physicians should pay more attention to these factors mentioned above when designing RCTs and evaluate whether these issues would influence the observed outcomes and to what degree. Since that not all women with PCOS are overweight or obese, the impact of diet independent of weight loss is of great clinical interest. This review only compared diet with minimal intervention, future studies should expand the research scope and make comparisons with other commonly used pharmacological and surgical treatments or explore the possibility of combining interventions.